Intel® Extension for PyTorch*: New Features on CPUs and GPUs

Intel® Extension for PyTorch* is a plug-in to PyTorch that provides further optimizations and features when run on Intel hardware, including CPUs and GPUs. Few code changes are needed to take full advantage of the available optimizations.

This session focuses on new and experimental features of Intel’s latest PyTorch extension, especially those supported in PyTorch 2.0 (released March 2023), which optimize models and runtime and further enable developers to take full advantage of Intel hardware capabilities.

Key topics covered:

• New features that provide additional optimizations using the extension’s back end for torch.compile(), codeless optimization, and fast BERT
• Graph capture to automatically generate a graph model from TorchScript trace and TorchDynamo
• HyperTune for quantizing models that have high accuracy loss when quantized using other methods
• New features in the extension’s launch script, including specifying numa nodes, use of P-cores only, and distributed training options
• For GPUs, distributed training using distributed data parallel (DDP) and Horovod*

The session includes a demo.

Skill level: Intermediate

Get the stand-alone version of Intel Extension of PyTorch from GitHub* or as part of the AI Frameworks and Tools.

• Get Started with Intel Extension for PyTorch
• Optimize PyTorch Models: Quantization Sample
• See All Code Samples